Detecting Deforestation With Sentinel-1 and ML

ISEF Category: Earth and Environmental Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Illegal logging can erase a forest patch before a satellite image update even arrives. Radar satellites can still “see” through clouds, which makes them useful in the tropics. Your project can test whether machine learning spots forest loss faster, and cleaner, than simple change rules. That gives you a real problem with real stakes.

What Is It?

This project asks you to detect forest loss from space, using radar data instead of normal photos. Sentinel-1 is a satellite system that sends out radar pulses and measures the signal that bounces back. Forests, bare soil, water, and cut land all return different signals, so change detection can reveal when a patch of forest suddenly looks less like forest.

Think of it like listening to a room instead of looking at it. A loud room and an empty room both hold the same space, but the echo changes a lot. Radar works in a similar way. A transformer is a machine learning model that looks for patterns in data and learns which changes match known deforestation alerts. Global Forest Watch alerts can act like your training labels, which are the examples the model learns from.

Why This Is a Good Topic

This is a strong science fair topic because you can define a clear signal, measure performance, and compare methods. You can test simple thresholds against a machine learning model, then ask which one catches deforestation faster or with fewer false alarms. The topic connects to forest conservation, carbon loss, and environmental monitoring, but you can still turn it into a focused, testable question. You also get real practice with geospatial data, model evaluation, and error analysis.

Research Questions

• How does a Sentinel-1 change detection threshold compare with a transformer model for identifying new deforestation alerts?
• What is the effect of different temporal window lengths on deforestation detection accuracy?
• Does training on one tropical region transfer well to another tropical region with different forest types?
• To what extent does cloud season affect the performance of optical versus radar-based deforestation detection?
• Which model features, such as backscatter change or texture change, best separate forest loss from normal seasonal variation?
• How does low-bandwidth compression affect the accuracy of an edge-deployed deforestation alert model?

Basic Materials

• Laptop with at least 16 GB RAM.
• Stable internet access for downloading satellite tiles and alert labels.
• External hard drive or cloud storage for geospatial files.
• Free GIS software such as QGIS.
• Python installed with geospatial and machine learning libraries.
• Jupyter Notebook for organizing analysis and plots.
• Public alert data from Global Forest Watch.
• Sentinel-1 scene access through a public data portal.

Advanced Materials

• Workstation or cloud compute account with GPU access.
• Large storage for Sentinel-1 time series and training data.
• Geospatial processing stack such as GDAL, Rasterio, GeoPandas, and xarray.
• Deep learning framework such as PyTorch or TensorFlow.
• High-resolution land cover or concession boundary data for validation.
• GIS layers for roads, protected areas, and land use context.
• Benchmark dataset of manually labeled deforestation events.
• Model compression tools for testing low-bandwidth deployment.

Software & Tools

• QGIS: Maps Sentinel-1 layers, checks geometry, and helps you inspect alert locations visually.
• Python: Processes satellite time series, builds features, and trains baseline and transformer models.
• Google Earth Engine: Helps you access and prefilter large remote sensing datasets without downloading everything first.
• ImageJ: Measures image-based change on exported rasters when you want a simple comparison workflow.
• PyTorch: Trains transformer models and supports model evaluation for classification tasks.

Experiment Steps

1. Define one tropical region and one deforestation event label source, so your project has a clear study area and ground truth.
2. Decide your comparison setup, such as simple change detection versus a transformer model, so you can test whether ML adds value.
3. Build the data pipeline that aligns Sentinel-1 scenes, alert labels, and map boundaries into one analysis table.
4. Choose features and input windows that capture real forest loss without overreacting to seasonal noise.
5. Plan your evaluation metrics, including precision, recall, false alarms, and detection delay.
6. Test a compressed version of the model, then compare accuracy and file size for low-bandwidth deployment.

Common Pitfalls

• Using misaligned satellite scenes, which makes a real forest change look like a model error.
• Training on alerts that overlap the test region, which leaks information and inflates accuracy.
• Ignoring seasonal flooding or soil moisture shifts, which can look like deforestation in radar data.
• Comparing models with different input windows, which makes the fairness of the test unclear.
• Skipping map-level error checks, which hides whether the model misses small clearings or edge cuts.

What Makes This Competitive

A strong version of this project does more than predict change. It explains when radar helps, when it fails, and why. You can stand out by comparing regions, testing transfer across landscapes, and measuring detection delay, not just final accuracy. A careful error analysis, plus a deployment test for low bandwidth settings, makes the work feel like a real monitoring tool rather than a class demo.

Project Variations

• Compare Sentinel-1 change detection with Sentinel-2 optical alerts in a cloud-prone tropical region.
• Test whether a model trained on one country transfers to a different forest frontier with different road patterns.
• Add roads, slope, and protected area boundaries as context features to see whether they reduce false alerts.

Learn More

• NASA ARSET Remote Sensing Training: Free lessons on satellite monitoring, including forest change, found through NASA ARSET course pages.
• Global Forest Watch: Public deforestation alert maps and datasets, available through the Global Forest Watch data portal.
• USGS EarthExplorer: Access to satellite imagery and land data, with guides for finding and downloading scenes.
• ESA Sentinel-1 Mission Pages: Background on radar imaging and Sentinel-1 data, available on the European Space Agency site.
• MIT OpenCourseWare Remote Sensing Courses: Free lecture materials on image analysis and Earth observation, found by searching MIT OpenCourseWare for remote sensing.
• PubMed: Search for review articles on satellite-based forest monitoring and machine learning validation.